Udbhav Bamba

ubamba98@gmail.com

I am an Applied Scientist II at Amazon Lab126, where I work at the intersection of large language models and low-latency on-device inference. My current work scales model compression to enable on-device and cloud-based GenAI deployment for Alexa+ — spanning dense-to-MoE conversion, quantization-friendly training, and compressing 200B+ parameter mixture-of-experts models. Previously at Amazon Ads, I built and productionized vision-language models for content understanding and moderation. Before Amazon, I focused on resource-efficient machine learning as a research intern at Mila – Quebec AI Institute (with David Rolnick and Gintare Karolina Dziugaite) and co-founded Transmute AI Research to foster a research culture among students at IIT (ISM) Dhanbad.

These days I am working on building better mixture-of-experts systems and helping LLMs reason efficiently under strict compute budgets. Beyond my research, I am an active competitor on Kaggle, where I hold the Competitions Master title (top 0.02%, peak rank 45th worldwide). Apart from work, I enjoy gaming, binge-watching shows, and playing badminton.

For updated details, please see my Google Scholar / LinkedIn pages.

news

Jun 01, 2026	Received the Silver Reviewer Award at ICML 2026, placing among the top reviewers.
May 01, 2026	Three papers accepted at ICML 2026 — XRPO (targeted exploration and exploitation for GRPO), DOT-MoE (differentiable optimal transport for MoEfication), and Reward Under Attack (robustness and hackability of process reward models).
Feb 26, 2026	S2D, on quantization-friendly conditioning of neural activations, accepted at CVPR 2026.
Jan 15, 2026	CRoPS, a training-free hallucination mitigation framework for vision-language models, accepted at TMLR.
Jul 01, 2025	Joined Amazon Lab126 as an Applied Scientist II, working on model compression for on-device and cloud GenAI deployment for Alexa+.

Selected publications

XRPO: Pushing the Limits of GRPO with Targeted Exploration and Exploitation

Udbhav Bamba, Minghao Fang, Yifan Yu, and 2 more authors

In International Conference on Machine Learning (ICML), 2026

Bib PDF

@inproceedings{bamba2025xrpo,
  title = {XRPO: Pushing the Limits of GRPO with Targeted Exploration and Exploitation},
  author = {Bamba, Udbhav and Fang, Minghao and Yu, Yifan and Zheng, Haizhong and Lai, Fan},
  booktitle = {International Conference on Machine Learning (ICML)},
  year = {2026},
  url = {https://arxiv.org/abs/2510.06672},
}

DOT-MoE: Differentiable Optimal Transport for MoEfication

Udbhav Bamba, Arnav Chavan, Aryamaan Thakur, and 2 more authors

In International Conference on Machine Learning (ICML), 2026

Bib PDF

@inproceedings{bamba2026dotmoe,
  title = {DOT-MoE: Differentiable Optimal Transport for MoEfication},
  author = {Bamba, Udbhav and Chavan, Arnav and Thakur, Aryamaan and Teig, Steve and Gupta, Deepak K.},
  booktitle = {International Conference on Machine Learning (ICML)},
  year = {2026},
  url = {https://arxiv.org/abs/2606.01666},
}

S2D: Selective Spectral Decay for Quantization-Friendly Conditioning of Neural Activations

Arnav Chavan, Nahush Lele, Udbhav Bamba, and 3 more authors

In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026

Bib PDF

@inproceedings{chavan2025s2d,
  title = {S2D: Selective Spectral Decay for Quantization-Friendly Conditioning of Neural Activations},
  author = {Chavan, Arnav and Lele, Nahush and Bamba, Udbhav and Dayal, Sankalp and Raghunathan, Aditi and Gupta, Deepak K.},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year = {2026},
  url = {https://arxiv.org/abs/2602.14432},
}

CRoPS: A Training-Free Hallucination Mitigation Framework for Vision-Language Models

Neeraj Anand, Samyak Jha, Udbhav Bamba, and 1 more author

Transactions on Machine Learning Research, 2026

Bib PDF

@article{anand2025crops,
  title = {CRoPS: A Training-Free Hallucination Mitigation Framework for Vision-Language Models},
  author = {Anand, Neeraj and Jha, Samyak and Bamba, Udbhav and Rahaman, Rahul},
  journal = {Transactions on Machine Learning Research},
  year = {2026},
  url = {https://openreview.net/forum?id=KQSoZDPVGX},
}

Reward Under Attack: Analyzing the Robustness and Hackability of Process Reward Models

Rishabh Tiwari, Aditya Tomar, Udbhav Bamba, and 5 more authors

In International Conference on Machine Learning (ICML), 2026

Bib PDF

@inproceedings{bamba2026reward,
  title = {Reward Under Attack: Analyzing the Robustness and Hackability of Process Reward Models},
  author = {Tiwari, Rishabh and Tomar, Aditya and Bamba, Udbhav and Maheswaran, Monishwaran and Yang, Heng and Mahoney, Michael W. and Keutzer, Kurt and Gholami, Amir},
  booktitle = {International Conference on Machine Learning (ICML)},
  year = {2026},
  url = {https://arxiv.org/abs/2603.06621},
}